Skip to content

Voice IDs

Kokoro ships eleven built-in English voices. Each one is a single .bin style-vector file under ai-models/tts/voices/. Pick by passing the voiceId (filename without .bin) to KokoroTtsRunner.SynthesizeAsync(text, voiceId).

All entries sourced verbatim from ai-models/tts/manifest.json and verified Apache-2.0.

Naming convention

Voice ids follow a two-letter prefix convention, optionally followed by an underscore and a speaker name:

   accent letter
   |
   v
   a  f  _bella
   ^  ^
   |  gender letter
   |
   first letter
Letter 1 (accent) Letter 2 (gender)
a — American English f — female
b — British English m — male

The blank af (with no underscore-suffix) is the default American-female blend — not a named speaker. Every other voice is one specific recorded speaker.

The full catalogue

American Female (af*)

Voice id Display name Notes
af American Female (default blend) Unnamed blend. Safe default.
af_bella American Female, Bella Candidate default voice.
af_nicole American Female, Nicole Candidate default voice.
af_sarah American Female, Sarah
af_sky American Female, Sky

American Male (am*)

Voice id Display name Notes
am_adam American Male, Adam
am_michael American Male, Michael Candidate default voice.

British Female (bf*)

Voice id Display name Notes
bf_emma British Female, Emma
bf_isabella British Female, Isabella

British Male (bm*)

Voice id Display name Notes
bm_george British Male, George
bm_lewis British Male, Lewis

On-disk shape

Each .bin file is 524 288 bytes (= 131 072 floats × 4 bytes/float) and represents a (512, 1, 256) tensor:

  • Leading dim 512 — "max token length" rows. The runner indexes by len(tokens) (the unwrapped token count, before pad wrapping) to pick the row used as the style vector for this utterance.
  • Middle dim 1 — batch dim (always 1).
  • Trailing dim 256 — the style vector dimensionality the Kokoro model expects.

The runner caches each voice's full tensor on first use; subsequent synth calls slice the relevant row without re-reading from disk. See KokoroTtsRunner.LoadVoiceStyleRow.

How to pick a voice

Pattern Try
Generic NPC, no strong persona af_bella or am_michael
Authority figure / officer / elder bm_george or bm_lewis
Mystic / scholar / narrator bf_emma or bf_isabella
Young or playful character af_sky or af_nicole
Quick prototyping / "just need a voice" af (the unnamed blend)

The pragmatic test: synthesise a few characteristic lines per voice ("Welcome, traveller, to Stormwall.", "I do not give my answers in daylight.", "The artifact lies beneath the lake.") and pick the one that fits the character.

You can mix voices freely across characters in the same scene — there's no extra cost beyond a one-time per-voice file read into the cache.

Per-character voice via templates

If you're authoring NPCs with the npc-dialogue.json template, set the voice via the voice.voiceId field:

{
  "voice": {
    "voiceId": "bf_emma",
    "speed": 0.9
  }
}

The speed modifier scales playback rate. 1.0 is natural; 0.8 for slower / more contemplative; 1.2 for hurried. Values too far from 1.0 produce audible artifacts.

What you don't get in v1.x

  • Non-English voices. Kokoro itself supports more, but Sauti is English-only per voice_ai_architecture.md § 10. The voices ship only the English style vectors.
  • Voice cloning. Kokoro is not a voice-cloning model. You cannot upload a few seconds of a recorded voice and have Kokoro mimic it.
  • Emotional control. Kokoro inflection is determined by the style vector and the punctuation in the input. There is no "emotion": "angry" parameter.

Cross-references